All reviews of published articles are made public. This includes manuscript files, peer review comments, author rebuttals and revised materials. Note: This was optional for articles submitted before 13 February 2023.
Peer reviewers are encouraged (but not required) to provide their names to the authors when submitting their peer review. If they agree to provide their name, then their personal profile page will reflect a public acknowledgment that they performed a review (even if the article is rejected). If the article is accepted, then reviewers who provided their name will be associated with the article itself.
Dear Authors,
One reviewer who has accepted the invitation to review has not yet submitted his/her review report. However your paper seems to address the relevant comments and criticisms. The manuscript is ready for publication.
Best wishes,
[# PeerJ Staff Note - this decision was reviewed and approved by Xiangjie Kong, a PeerJ Section Editor covering this Section #]
Dear Authors,
Two of the previous reviewers did not respond to the invitation for reviewing the revised paper. Although one reviewer accepts your paper, one reviewer suggests minor revision. We encourage you to address the concerns and criticisms of Reviewer 1 and resubmit your paper once you have updated it accordingly.
Best wishes,
**PeerJ Staff Note:** It is PeerJ policy that additional references suggested during the peer-review process should only be included if the authors agree that they are relevant and useful.
The revised manuscript proposes a Mixup-based augmentation method tailored for long-tailed image classification, combining instance-based and class-based sampling with a decay mixing strategy. Experiments on multiple benchmarks show consistent improvements over existing methods. The approach is intuitive, effective, and the revisions have addressed previous concerns.
However, several minor issues still require attention and improvement:
1.To more comprehensively evaluate the performance of the proposed method, it is recommended to include comparisons with representative methods published in the past two years to enhance the persuasiveness of the results.
2.To better demonstrate the practical value of the proposed method, please provide an analysis of its computational complexity and runtime efficiency.
3.Dempster-Shafer (DS) evidence theory has shown promising potential in addressing data imbalance problems (e.g., Doi: 10.1109/TSMC.2022.3162258; 0.1016/j.inffus.2024.102736; 10.1007/978-3-031-17801-6_8). The authors may consider briefly discussing its possible relevance to the current study.
N/A
N/A
N/A
Paper is clearly written, references are sufficient, logical professional structure.
Material has been arrange in a good style
The experimental design is now much better.
Conclusion is good and valid - as far as I can see
Paper has been revised in accordance to my and the other reviewers comments.
Dear Authors,
Thank you for the submission. The reviewers’ comments are now available. It is not suggested that your article be published in its current format. We do, however, advise you to revise the paper in light of the reviewers’ concerns with respect to basic reporting, experimental design, validity of of the findings, and additional comments before resubmitting it.
By the way, explanation of the equations should be checked. All variables should be written in italics as in the equations. Their definitions and boundaries should be defined. Necessary references should also be given. Many of the equations are part of the related sentences. Attention is needed for correct sentence formation.
Best wishes,
**PeerJ Staff Note:** Please ensure that all review and editorial comments are addressed in a response letter and that any edits or clarifications mentioned in the letter are also inserted into the revised manuscript where appropriate.
**Language Note:** The review process has identified that the English language must be improved. PeerJ can provide language editing services - please contact us at [email protected] for pricing (be sure to provide your manuscript number and title). Alternatively, you should make your own arrangements to improve the language quality and provide details in your response letter. – PeerJ Staff
This paper proposed a Mixup-based data augmentation method to deal with long-tailed image classification problems. It utilizes two sampling techniques to independently sample and mix the samples. The following comments are some suggestions:
1. The presentation of the introduction is weak, and the structure is inappropriate. It is difficult to grasp the contribution of the proposed method.
2. The motivations or remaining challenges could be clearer. Please give more details and discuss the key problems solved in this paper.
3. The authors proposed a mixup-based improved data augmentation method. It is suggested to provide a detailed explanation of the limitations of the original mixup approach and clearly describe the specific aspects where improvements were made.
4.It is suggested to conduct a comprehensive review of recent literature on long-tailed classification.
5. The explanation provided below Equation (5) appears to be incorrect; it should involve x and y. It is recommended to thoroughly review and verify all the formulas to ensure their accuracy.
6. The authors should compare their method with related methods proposed in the past two years.
-
-
-
The paper addresses the challenge of image classification in long-tailed distributions, where few classes dominate the training data while many classes have limited representation. It introduces a novel Mixup-based data augmentation technique that combines two sampling mechanisms: instance-based sampling and class-based sampling, to create a more balanced sample distribution. The method also incorporates a decay mixing strategy to adjust sample mixing weights dynamically during training, preventing overfitting. Experimental results show that the proposed method outperforms other advanced Mixup-based techniques on long-tailed datasets like CIFAR-10, CIFAR-100, CINIC-10, Tiny-ImageNet, and Fashion-Mnist.
Detailed comments:
1. The proposed Mixup-based data augmentation technique is innovative and effectively addresses the long-tailed distribution problem in image classification. However, it would be helpful to more explicitly compare the novelty of this approach with existing methods in terms of theoretical foundations and practical impact to highlight its unique contribution.
2. The experimental validation across multiple long-tailed datasets demonstrates the method’s effectiveness. However, additional details on the exact distribution of training samples for each dataset (i.e., the extent of imbalance) would help contextualize the results and clarify the real-world applicability of the approach.
3. While the proposed method shows improved performance, the paper could benefit from a deeper analysis of the impact of the decay mixing strategy, especially with respect to different training phases. For example, a comparison of results when this strategy is applied at different intervals during training would be valuable.
4. The method performs well on the tested datasets, but its generalizability to other domains (e.g., object detection, segmentation) or more extreme long-tail imbalances is not discussed. A section on potential extensions of the method to other tasks or domains would enhance the overall impact and vision of the work.
-
-
-
The English is clear, at least from my perspective.
Literature review, Abstract, and Introduction are fine.
The paper is written in a good style, the structure is logical, figures and tables are clear. Here I suggest extending the captions a bit more, e.g., in Tables 1-7, the numbers 10, 50 100 should be explained. These numbers are clear from the text, but it would help to explain the numbers in the first table caption.
In the paper, an algorithm is motivated, presented, and evaluated - we have no formal results, say on a mathematical level of description.
In section "Proposed Methods," the Eq. (3) and (7) must be checked/ corrected.
Page 4, bottom line " ... of class. $ I guess here a $j$ is missing
$B_C$ and $B_I$ must be fine-tuned in a better way, not in the context of $q$ (see page 5)
Page 5, 2 lines to explain Eq (5) must be corrected - a copy and paste failure.
All data sets used in the study are constructed so that they are imbalanced. From my perspective experiments on real imbalanced data sets are not used in the experiments. In case of down-sampling a big class to a smaller one reduces the number of samples but does not change its distribution.
Thus, I suggest including experiments with data of real imbalanced data.
Here I miss a rigorous statistical evaluation using statistical testing. To draw conclusions based on point estimates, such as averages or medians, is too weak.
- The presentation of this paper is very fluent, clear, and detailed.
- Some confusing equations are:
- Eq.3, should $(n_j)^k$ be $(n_k)^q$?
- There might be a formatting mistake with $ezk$ in Eq.7.
- The experimental results do not include standard deviations of the significance test, making it less convincing.
- Since the problem is about the unbalanced dataset, evaluation metrics such as recall and F1 score are very meaningful criteria and should be reported.
- Some of the latest baselines are not discussed or compared, especially DBN-Mix(https://doi.org/10.1016/j.patcog.2023.110107), which is of high similarity with this paper.
- Experiments discussing the effect of using linear decay are very insightful and meaningful
- Data is well-specified and code is provided, making it convenient to replicate.
- Conclusions are well supported by empirical evidence.
- The proposed method is similar to BNN at first sight, but the authors explain their differences and demonstrate the superiority of the proposed method. Also, the authors use a more effective decay (linear), which is also different and advantageous. (It would be better if the architectural difference is stated earlier in the paper.) I do appreciate the effort to demonstrate the superiority of using linear decay.
- A similar design is also adopted by D-Mixup (https://doi.org/10.1109/TMM.2022.3181789) and DBN-Mix, with the latter seemingly more effective than the proposed method. Given their existence, I suggest making some enhancements to the method to improve its performance or demonstrating the advantages of the proposed method from other aspects (such as time and memory) in the paper.
All text and materials provided via this peer-review history page are made available under a Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.